13
Tentative Conclusions
¥ For individual subject thresholded activation maps:
H Use GLSQ/REML estimation, especially for slow block design experiments at 3+ Tesla
H Be aware that there may be many false negatives
o i.e., false acceptances of the null hypothesis
o am looking into an FDR-like procedure for estimating the false negative rate, similar to how FDR estimates the false positive rate

¥ For group maps using ANOVA (or similar statistics):
H Differences between OLSQ and GLSQ estimation are small

¥ Recommendations:
H DonÕt need to re-visit group activation conclusions!
H Use 3dREMLfit as a near drop-in replacment for 3dDeconvolve for future work
oA little extra CPU time (usually from 1..3 times as long)